I’ve seen the same. Also it’s very difficult to get the instructions and the attached PDF’s to work together well. I have spent hours on this in the Assistant API. I could really use some good documentation in this area.
There is problem that you need to be aware of: Files uploaded via chat have a 3 hour window to be accessed by GPTs. You should upload new documents when the time comes.
And should note that GPTs currently avoid repeating or reproducing text from documents you provide. To avoid creating pirated output. Even if the document is yours.
As many of the lawsuits against OpenAI have done in practice, this can happen when anyone can input a document and have the AI speak its contents. Although the charges were ultimately dismissed, But what happened caused OpenAI to lose its reputation and choose to no longer use any external data to train its AI or enter the system.
And PDF is a document that is not suitable for use with AI. It is recommended that using .text will be more effective. You can also write a prompt to do certain tasks in the document before loading. To make work easier
Importantly, using GPTs rips out the RAG manual as it doesn’t tell you anything about AI behavior affecting the document.
Hello Chieffy99, very interesting answer about problems with pdf documents that I have also encountered. In your experience what would be the best approach with scientific PDF paper with many equations?. I was trying diffrent options even converting to text but is not possible to working with citations and on the specific chapters. Chat is trying always make the summary that partialy is his hallucinations , I need help because I starting loosing my convince about this AI tool ;).
I did simple test , I gave him text or pdf document ( no more like 8 pages) and asked him first to print out abstract as it was in the orginal source and next do the same for chapter 1, 2, 3, informed him that we doing this for test of his capabilities. Next asked him to list equations in the table. For my suprice this chat is not capable to accomplish this task without halucination and many errors and repeats. From my observation abstract can more or less properly print out but chapters and equations are completly wrong and he is always trying add some text that does not exisit in the orginal paper. This is realy dangerous for profesional work. You could imagine when you developing some approach to land on the Mars with ChatGPT ;))). At this moment this approach is too dangerous for proffesional use. Im realy suprice that Chat has problem with that basic task like working with text.
For everybody who wants try my pdf papers to test chatgpt capabilities as the scientific tool and prepare single question matrix of error, please give me info I will send. We can quantified those problems in the matrix.
How you put table in .text file ?
Is AI use data analysis ?
Why i say PDF is unsuitable for knowledge.?
Even if you use a simple table with just a few columns or rows, there are limits to data can access. As in the picture I presented,
AI has no information about The sample was blind, although it was subjected to repeated testing. Until I used a screenshot to confirm. Therefore, there will be errors when using the data further. If it is a .text file such as a note pad, most of everything that is entered will not have this kind of problem. In addition, the behavior of GPTs with ChatGPT4 also uses tools. to manage documents differently. Even though GPT4 is smarter, if there is no guidance or clear instructions Calculations can be made in other directions.
I mean Ive tried this and Ive still been told that it is unable to read it via code interpreter did you do anything prior to this such as changing file format or was it just a PDF.
As of last update, it can no longer provide readings or interpretations based on images, such as those containing text or formulas from documents, books, or screenshots. Due to copyright and other restrictions.
Depend on what model. GPTs have a stubborn behavior in using documentation tools. Lack of flexibility. GPT-4 is smarter in this aspect. It accepting files simultaneously and use code interpretation file by file and summarize the results and this tool will not be used in Knowledge.
Is your prompt in this case read or repeat? Due to behavior to prevent copyright problems Try to summarize it and you will see more results. Or you might use it while doing back-end work.
, I need help because I starting loosing my convince about this AI tool ;).
